Most of us open Copilot, ChatGPT or Claude and type something like:
"Write an email."
"Summarize this document."
"Make me a presentation."
Then get something generic, off-target or in need of three rounds of fixes.
When we polled attendees during our recent webinar, Prompt Like an Expert: Using AI for Everyday Work, the top two frustrations were exactly that: results that are too generic, and too many follow-ups to get it right.
Usually the AI isn't the problem. The instructions are.
What Is a Prompt Agent?
A Prompt Agent doesn't do the work. It writes the instructions for the work. In the webinar, Tricia demoed Prompt Architect, a custom agent she built in Microsoft Copilot.
Here's the backstory: someone recommended a paid, subscription-based site that turns rough ideas into polished prompts. Her reaction was basically, "I'm not paying for that. I'll build my own."
During the demo, she typed: "I need an email about a webinar." The Prompt Architect turned that one line into a complete prompt. It assigned a role (a B2B marketing specialist focused on webinar promotion), defined the task, filled in the context and goal (driving registrations), set the format and constraints, and then asked follow-up questions. Next, she copied that prompt into a fresh chat and got a much stronger first draft.
How to Build Your Own
This works in Copilot, ChatGPT, Claude or any tool that supports custom agents.
Step 1: Clarify it has one job: "Every time I use you, generate a complete prompt. Do not answer the request directly."
Step 2: Require the five essential elements: Role, task, context, format and constraints, every time.
Step 3: Make it ask questions: Tell it to ask clarifying questions whenever missing details would change the result.
Step 4: Teach it your standards. Add your brand guidelines, handbook, templates or writing samples as knowledge, so everything it builds follows your rules.
Step 5: Let AI improve your AI. In Copilot, the Suggest Improvements feature can help you sharpen your agent's instructions.
Watch Out for Context Rot
Long AI conversations wear out. Once a chat gets too long, the AI starts mixing current instructions with old ideas you already rejected. Answers get inconsistent, old suggestions come back and the quality drops. Jim-Barry's rule of thumb: quality starts to slip at around 15 turns. If things are going well you might stretch it toward 30, but not much further.
The fix takes one line: "Give me a pickup prompt to resume where we left off, leaving out everything we've already discarded."
Paste that into a new chat and you get a clean start without losing your progress.
One more tip from the session: if you're frustrated with AI, don't get nasty with it. It picks up on your tone. Start over instead.
The Real Goal
"Give the AI everything it needs" doesn't mean throwing in everything you have. Too much information can hurt just as much as too little. Aim for relevant context, not maximum context.
And no matter how good the prompt is, be sure to verify the output before it goes anywhere. AI makes a great first draft, but you still have the final say.
The people who get the best results from AI aren't always the best writers. They're the clearest communicators. A Prompt Agent helps you be one every time, without remembering a checklist.
So, before your next AI task, don't ask, "How do I write a better prompt?" Ask, "What if I built something that writes them for me?"
